magenta-retry / jam_worker.py
thecollabagepatch's picture
reseed splice context fix
bf8ae4c
raw
history blame
24.1 kB
# jam_worker.py - SIMPLE FIX VERSION
import threading, time, base64, io, uuid
from dataclasses import dataclass, field
import numpy as np
import soundfile as sf
from magenta_rt import audio as au
from threading import RLock
from utils import (
match_loudness_to_reference, stitch_generated, hard_trim_seconds,
apply_micro_fades, make_bar_aligned_context, take_bar_aligned_tail,
resample_and_snap, wav_bytes_base64
)
@dataclass
class JamParams:
bpm: float
beats_per_bar: int
bars_per_chunk: int
target_sr: int
loudness_mode: str = "auto"
headroom_db: float = 1.0
style_vec: np.ndarray | None = None
ref_loop: any = None
combined_loop: any = None
guidance_weight: float = 1.1
temperature: float = 1.1
topk: int = 40
@dataclass
class JamChunk:
index: int
audio_base64: str
metadata: dict
class JamWorker(threading.Thread):
def __init__(self, mrt, params: JamParams):
super().__init__(daemon=True)
self.mrt = mrt
self.params = params
self.state = mrt.init_state()
# βœ… init synchronization + placeholders FIRST
self._lock = threading.Lock()
self._original_context_tokens = None # so hasattr checks are cheap/clear
if params.combined_loop is not None:
self._setup_context_from_combined_loop()
self.idx = 0
self.outbox: list[JamChunk] = []
self._stop_event = threading.Event()
self._stream = None
self._next_emit_start = 0
# NEW: Track delivery state
self._last_delivered_index = 0
self._max_buffer_ahead = 5
# Timing info
self.last_chunk_started_at = None
self.last_chunk_completed_at = None
def _setup_context_from_combined_loop(self):
"""Set up MRT context tokens from the combined loop audio"""
try:
from utils import make_bar_aligned_context, take_bar_aligned_tail
codec_fps = float(self.mrt.codec.frame_rate)
ctx_seconds = float(self.mrt.config.context_length_frames) / codec_fps
loop_for_context = take_bar_aligned_tail(
self.params.combined_loop,
self.params.bpm,
self.params.beats_per_bar,
ctx_seconds
)
tokens_full = self.mrt.codec.encode(loop_for_context).astype(np.int32)
tokens = tokens_full[:, :self.mrt.config.decoder_codec_rvq_depth]
context_tokens = make_bar_aligned_context(
tokens,
bpm=self.params.bpm,
fps=float(self.mrt.codec.frame_rate), # keep fractional fps
ctx_frames=self.mrt.config.context_length_frames,
beats_per_bar=self.params.beats_per_bar
)
# Install fresh context
self.state.context_tokens = context_tokens
print(f"βœ… JamWorker: Set up fresh context from combined loop")
# NEW: keep a copy of the *original* context tokens for future splice-reseed
# (guard so we only set this once, at jam start)
with self._lock:
if not hasattr(self, "_original_context_tokens") or self._original_context_tokens is None:
self._original_context_tokens = np.copy(context_tokens) # shape: [T, depth]
except Exception as e:
print(f"❌ Failed to setup context from combined loop: {e}")
def stop(self):
self._stop_event.set()
def update_knobs(self, *, guidance_weight=None, temperature=None, topk=None):
with self._lock:
if guidance_weight is not None: self.params.guidance_weight = float(guidance_weight)
if temperature is not None: self.params.temperature = float(temperature)
if topk is not None: self.params.topk = int(topk)
def get_next_chunk(self) -> JamChunk | None:
"""Get the next sequential chunk (blocks/waits if not ready)"""
target_index = self._last_delivered_index + 1
# Wait for the target chunk to be ready (with timeout)
max_wait = 30.0 # seconds
start_time = time.time()
while time.time() - start_time < max_wait and not self._stop_event.is_set():
with self._lock:
# Look for the exact chunk we need
for chunk in self.outbox:
if chunk.index == target_index:
self._last_delivered_index = target_index
print(f"πŸ“¦ Delivered chunk {target_index}")
return chunk
# Not ready yet, wait a bit
time.sleep(0.1)
# Timeout or stopped
return None
def mark_chunk_consumed(self, chunk_index: int):
"""Mark a chunk as consumed by the frontend"""
with self._lock:
self._last_delivered_index = max(self._last_delivered_index, chunk_index)
print(f"βœ… Chunk {chunk_index} consumed")
def _should_generate_next_chunk(self) -> bool:
"""Check if we should generate the next chunk (don't get too far ahead)"""
with self._lock:
# Don't generate if we're already too far ahead
if self.idx > self._last_delivered_index + self._max_buffer_ahead:
return False
return True
def _seconds_per_bar(self) -> float:
return self.params.beats_per_bar * (60.0 / self.params.bpm)
def _snap_and_encode(self, y, seconds, target_sr, bars):
cur_sr = int(self.mrt.sample_rate)
x = y.samples if y.samples.ndim == 2 else y.samples[:, None]
x = resample_and_snap(x, cur_sr=cur_sr, target_sr=target_sr, seconds=seconds)
b64, total_samples, channels = wav_bytes_base64(x, target_sr)
meta = {
"bpm": int(round(self.params.bpm)),
"bars": int(bars),
"beats_per_bar": int(self.params.beats_per_bar),
"sample_rate": int(target_sr),
"channels": channels,
"total_samples": total_samples,
"seconds_per_bar": self._seconds_per_bar(),
"loop_duration_seconds": bars * self._seconds_per_bar(),
"guidance_weight": self.params.guidance_weight,
"temperature": self.params.temperature,
"topk": self.params.topk,
}
return b64, meta
def _append_model_chunk_to_stream(self, wav):
"""Incrementally append a model chunk with equal-power crossfade."""
xfade_s = float(self.mrt.config.crossfade_length)
sr = int(self.mrt.sample_rate)
xfade_n = int(round(xfade_s * sr))
s = wav.samples if wav.samples.ndim == 2 else wav.samples[:, None]
if getattr(self, "_stream", None) is None:
# First chunk: drop model pre-roll (xfade head)
if s.shape[0] > xfade_n:
self._stream = s[xfade_n:].astype(np.float32, copy=True)
else:
self._stream = np.zeros((0, s.shape[1]), dtype=np.float32)
self._next_emit_start = 0 # pointer into _stream (model SR samples)
return
# Crossfade last xfade_n samples of _stream with head of new s
if s.shape[0] <= xfade_n or self._stream.shape[0] < xfade_n:
# Degenerate safeguard
self._stream = np.concatenate([self._stream, s], axis=0)
return
tail = self._stream[-xfade_n:]
head = s[:xfade_n]
# Equal-power envelopes
t = np.linspace(0, np.pi/2, xfade_n, endpoint=False, dtype=np.float32)[:, None]
eq_in, eq_out = np.sin(t), np.cos(t)
mixed = tail * eq_out + head * eq_in
self._stream = np.concatenate([self._stream[:-xfade_n], mixed, s[xfade_n:]], axis=0)
def reseed_from_waveform(self, wav):
# 1) Re-init state
new_state = self.mrt.init_state()
# 2) Build bar-aligned context tokens from provided audio
codec_fps = float(self.mrt.codec.frame_rate)
ctx_seconds = float(self.mrt.config.context_length_frames) / codec_fps
from utils import take_bar_aligned_tail, make_bar_aligned_context
tail = take_bar_aligned_tail(wav, self.params.bpm, self.params.beats_per_bar, ctx_seconds)
tokens_full = self.mrt.codec.encode(tail).astype(np.int32)
tokens = tokens_full[:, :self.mrt.config.decoder_codec_rvq_depth]
context_tokens = make_bar_aligned_context(tokens,
bpm=self.params.bpm, fps=float(self.mrt.codec.frame_rate),
ctx_frames=self.mrt.config.context_length_frames,
beats_per_bar=self.params.beats_per_bar
)
new_state.context_tokens = context_tokens
self.state = new_state
self._prepare_stream_for_reseed_handoff()
def _frames_per_bar(self) -> int:
# codec frame-rate (frames/s) -> frames per musical bar
fps = float(self.mrt.codec.frame_rate)
sec_per_bar = (60.0 / float(self.params.bpm)) * float(self.params.beats_per_bar)
return int(round(fps * sec_per_bar))
def _ctx_frames(self) -> int:
# how many codec frames fit in the model’s conditioning window
return int(self.mrt.config.context_length_frames)
def _make_recent_tokens_from_wave(self, wav) -> np.ndarray:
"""
Encode waveform and produce a BAR-ALIGNED context token window.
"""
tokens_full = self.mrt.codec.encode(wav).astype(np.int32) # [T, rvq_total]
tokens = tokens_full[:, :self.mrt.config.decoder_codec_rvq_depth]
from utils import make_bar_aligned_context
ctx = make_bar_aligned_context(
tokens,
bpm=self.params.bpm,
fps=float(self.mrt.codec.frame_rate), # keep fractional fps
ctx_frames=self.mrt.config.context_length_frames,
beats_per_bar=self.params.beats_per_bar
)
return ctx
def _bar_aligned_tail(self, tokens: np.ndarray, bars: float) -> np.ndarray:
"""
Take a tail slice that is an integer number of codec frames corresponding to `bars`.
We round to nearest frame to stay phase-consistent with codec grid.
"""
frames_per_bar = self._frames_per_bar()
want = max(frames_per_bar * int(round(bars)), 0)
if want == 0:
return tokens[:0] # empty
if tokens.shape[0] <= want:
return tokens
return tokens[-want:]
def _splice_context(self, original_tokens: np.ndarray, recent_tokens: np.ndarray,
anchor_bars: float) -> np.ndarray:
import math
ctx_frames = self._ctx_frames()
depth = original_tokens.shape[1]
frames_per_bar = self._frames_per_bar()
# 1) Anchor tail (whole bars)
anchor = self._bar_aligned_tail(original_tokens, math.floor(anchor_bars))
# 2) Fill remainder with recent (prefer whole bars)
a = anchor.shape[0]
remain = max(ctx_frames - a, 0)
recent = recent_tokens[:0]
used_recent = 0 # frames taken from the END of recent_tokens
if remain > 0:
bars_fit = remain // frames_per_bar
if bars_fit >= 1:
want_recent_frames = int(bars_fit * frames_per_bar)
used_recent = min(want_recent_frames, recent_tokens.shape[0])
recent = recent_tokens[-used_recent:] if used_recent > 0 else recent_tokens[:0]
else:
used_recent = min(remain, recent_tokens.shape[0])
recent = recent_tokens[-used_recent:] if used_recent > 0 else recent_tokens[:0]
# 3) Concat in order [anchor, recent]
if anchor.size or recent.size:
out = np.concatenate([anchor, recent], axis=0)
else:
# fallback: just take the last ctx window from recent
out = recent_tokens[-ctx_frames:]
# 4) Trim if we overshot
if out.shape[0] > ctx_frames:
out = out[-ctx_frames:]
# 5) Snap the **END** to the nearest LOWER bar boundary
if frames_per_bar > 0:
max_bar_aligned = (out.shape[0] // frames_per_bar) * frames_per_bar
else:
max_bar_aligned = out.shape[0]
if max_bar_aligned > 0 and out.shape[0] != max_bar_aligned:
out = out[-max_bar_aligned:]
# 6) Left-fill to reach ctx_frames **without moving the END**
deficit = ctx_frames - out.shape[0]
if deficit > 0:
left_parts = []
# Prefer frames immediately BEFORE the region we used from 'recent_tokens'
if used_recent < recent_tokens.shape[0]:
take = min(deficit, recent_tokens.shape[0] - used_recent)
if used_recent > 0:
left_parts.append(recent_tokens[-(used_recent + take) : -used_recent])
else:
left_parts.append(recent_tokens[-take:])
# Then take frames immediately BEFORE the 'anchor' in original_tokens
if sum(p.shape[0] for p in left_parts) < deficit and anchor.shape[0] > 0:
need = deficit - sum(p.shape[0] for p in left_parts)
a_len = anchor.shape[0]
avail = max(original_tokens.shape[0] - a_len, 0)
take2 = min(need, avail)
if take2 > 0:
left_parts.append(original_tokens[-(a_len + take2) : -a_len])
# Still short? tile from what's available
have = sum(p.shape[0] for p in left_parts)
if have < deficit:
base = out if out.shape[0] > 0 else (recent_tokens if recent_tokens.shape[0] > 0 else original_tokens)
reps = int(np.ceil((deficit - have) / max(1, base.shape[0])))
left_parts.append(np.tile(base, (reps, 1))[: (deficit - have)])
left = np.concatenate(left_parts, axis=0)
out = np.concatenate([left[-deficit:], out], axis=0)
# 7) Final guard to exact length
if out.shape[0] > ctx_frames:
out = out[-ctx_frames:]
elif out.shape[0] < ctx_frames:
reps = int(np.ceil(ctx_frames / max(1, out.shape[0])))
out = np.tile(out, (reps, 1))[-ctx_frames:]
# 8) Depth guard
if out.shape[1] != depth:
out = out[:, :depth]
return out
def _realign_emit_pointer_to_bar(self, sr_model: int):
"""Advance _next_emit_start to the next bar boundary in model-sample space."""
bar_samps = int(round(self._seconds_per_bar() * sr_model))
if bar_samps <= 0:
return
phase = self._next_emit_start % bar_samps
if phase != 0:
self._next_emit_start += (bar_samps - phase)
def _prepare_stream_for_reseed_handoff(self):
# OLD: keep crossfade tail -> causes phase offset
# sr = int(self.mrt.sample_rate)
# xfade_s = float(self.mrt.config.crossfade_length)
# xfade_n = int(round(xfade_s * sr))
# if getattr(self, "_stream", None) is not None and self._stream.shape[0] > 0:
# tail = self._stream[-xfade_n:] if self._stream.shape[0] > xfade_n else self._stream
# self._stream = tail.copy()
# else:
# self._stream = None
# NEW: throw away the tail completely; start fresh
self._stream = None
self._next_emit_start = 0
self._needs_bar_realign = True
def reseed_splice(self, recent_wav, anchor_bars: float):
"""
Token-splice reseed:
- original = the context we captured when the jam started
- recent = tokens from the provided recent waveform (usually Swift-combined mix)
- anchor_bars controls how much of the original vibe we re-inject
"""
with self._lock:
if not hasattr(self, "_original_context_tokens") or self._original_context_tokens is None:
# Fallback: if we somehow don’t have originals, treat current as originals
self._original_context_tokens = np.copy(self.state.context_tokens)
recent_tokens = self._make_recent_tokens_from_wave(recent_wav) # [T, depth]
new_ctx = self._splice_context(self._original_context_tokens, recent_tokens, anchor_bars)
# install the new context window
new_state = self.mrt.init_state()
new_state.context_tokens = new_ctx
self.state = new_state
self._prepare_stream_for_reseed_handoff()
# optional: ask streamer to drop an intro crossfade worth of audio right after reseed
self._pending_drop_intro_bars = getattr(self, "_pending_drop_intro_bars", 0) + 1
def run(self):
"""Main worker loop β€” generate into a continuous stream, then emit bar-aligned slices."""
spb = self._seconds_per_bar() # seconds per bar
chunk_secs = self.params.bars_per_chunk * spb
xfade = float(self.mrt.config.crossfade_length) # seconds
sr = int(self.mrt.sample_rate)
chunk_samps = int(round(chunk_secs * sr))
def _need(first_chunk_extra=False):
"""How many more samples we still need in the stream to emit next slice."""
have = 0 if getattr(self, "_stream", None) is None else self._stream.shape[0] - getattr(self, "_next_emit_start", 0)
want = chunk_samps
if first_chunk_extra:
# reserve two bars extra so first-chunk onset alignment has material
want += int(round(2 * spb * sr))
return max(0, want - have)
def _mono_env(x: np.ndarray, sr: int, win_ms: float = 10.0) -> np.ndarray:
if x.ndim == 2: x = x.mean(axis=1)
x = np.abs(x).astype(np.float32)
w = max(1, int(round(win_ms * 1e-3 * sr)))
if w > 1:
kern = np.ones(w, dtype=np.float32) / float(w)
x = np.convolve(x, kern, mode="same")
d = np.diff(x, prepend=x[:1])
d[d < 0] = 0.0
return d
def _estimate_first_offset_samples(ref_loop_wav, gen_head_wav, sr: int, spb: float) -> int:
"""Tempo-aware first-downbeat offset (positive => model late)."""
try:
max_ms = int(max(160.0, min(0.25 * spb * 1000.0, 450.0)))
ref = ref_loop_wav if ref_loop_wav.sample_rate == sr else ref_loop_wav.resample(sr)
n_bar = int(round(spb * sr))
ref_tail = ref.samples[-n_bar:, :] if ref.samples.shape[0] >= n_bar else ref.samples
gen_head = gen_head_wav.samples[: int(2 * n_bar), :]
if ref_tail.size == 0 or gen_head.size == 0:
return 0
# envelopes + z-score
import numpy as np
def _z(a):
m, s = float(a.mean()), float(a.std() or 1.0); return (a - m) / s
e_ref = _z(_mono_env(ref_tail, sr)).astype(np.float32)
e_gen = _z(_mono_env(gen_head, sr)).astype(np.float32)
# upsample x4 for finer lag
def _upsample(a, r=4):
n = len(a); grid = np.arange(n, dtype=np.float32)
fine = np.linspace(0, n - 1, num=n * r, dtype=np.float32)
return np.interp(fine, grid, a).astype(np.float32)
up = 4
e_ref_u, e_gen_u = _upsample(e_ref, up), _upsample(e_gen, up)
max_lag_u = int(round((max_ms / 1000.0) * sr * up))
seg = min(len(e_ref_u), len(e_gen_u))
e_ref_u = e_ref_u[-seg:]
pad = np.zeros(max_lag_u, dtype=np.float32)
e_gen_u_pad = np.concatenate([pad, e_gen_u, pad])
best_lag_u, best_score = 0, -1e9
for lag_u in range(-max_lag_u, max_lag_u + 1):
start = max_lag_u + lag_u
b = e_gen_u_pad[start : start + seg]
denom = (np.linalg.norm(e_ref_u) * np.linalg.norm(b)) or 1.0
score = float(np.dot(e_ref_u, b) / denom)
if score > best_score:
best_score, best_lag_u = score, lag_u
return int(round(best_lag_u / up))
except Exception:
return 0
print("πŸš€ JamWorker started (bar-aligned streaming)…")
while not self._stop_event.is_set():
if not self._should_generate_next_chunk():
time.sleep(0.25)
continue
# 1) Generate until we have enough material in the stream
need = _need(first_chunk_extra=(self.idx == 0))
while need > 0 and not self._stop_event.is_set():
with self._lock:
style_vec = self.params.style_vec
self.mrt.guidance_weight = float(self.params.guidance_weight)
self.mrt.temperature = float(self.params.temperature)
self.mrt.topk = int(self.params.topk)
wav, self.state = self.mrt.generate_chunk(state=self.state, style=style_vec)
self._append_model_chunk_to_stream(wav) # equal-power xfade into a persistent stream
need = _need(first_chunk_extra=(self.idx == 0))
if self._stop_event.is_set():
break
# 2) One-time: align the emit pointer to the groove
if self.idx == 0 and self.params.combined_loop is not None:
# Compare ref tail vs the head of what we're about to emit
head_len = min(self._stream.shape[0] - self._next_emit_start, int(round(2 * spb * sr)))
seg = self._stream[self._next_emit_start : self._next_emit_start + head_len]
gen_head = au.Waveform(seg.astype(np.float32, copy=False), sr).as_stereo()
offs = _estimate_first_offset_samples(self.params.combined_loop, gen_head, sr, spb)
if offs != 0:
# positive => model late: skip some samples; negative => model early: "rewind" by padding
self._next_emit_start = max(0, self._next_emit_start + offs)
print(f"🎯 First-chunk offset compensation: {offs/sr:+.3f}s")
# snap to next bar boundary
self._realign_emit_pointer_to_bar(sr)
# 3) Emit exactly bars_per_chunk Γ— spb from the stream
start = self._next_emit_start
end = start + chunk_samps
if end > self._stream.shape[0]:
# shouldn't happen often; generate a bit more and loop
continue
slice_ = self._stream[start:end]
self._next_emit_start = end
y = au.Waveform(slice_.astype(np.float32, copy=False), sr).as_stereo()
# 4) Post-processing / loudness
if self.idx == 0 and self.params.ref_loop is not None:
y, _ = match_loudness_to_reference(
self.params.ref_loop, y,
method=self.params.loudness_mode,
headroom_db=self.params.headroom_db
)
else:
apply_micro_fades(y, 3)
# 5) Resample + exact-length snap + encode
b64, meta = self._snap_and_encode(
y, seconds=chunk_secs, target_sr=self.params.target_sr, bars=self.params.bars_per_chunk
)
meta["xfade_seconds"] = xfade
# 6) Publish
with self._lock:
self.idx += 1
self.outbox.append(JamChunk(index=self.idx, audio_base64=b64, metadata=meta))
if len(self.outbox) > 10:
cutoff = self._last_delivered_index - 5
self.outbox = [ch for ch in self.outbox if ch.index > cutoff]
print(f"βœ… Completed chunk {self.idx}")
print("πŸ›‘ JamWorker stopped")